The Promise and Challenges of Quantum Computing Potential Energy Surfaces

Abstract

Machine learned interatomic potentials (MLIPs) are now accurate enough that their errors are often smaller than those of the density functional theory data used to train them. Further progress requires more accurate reference data, and quantum computers have been proposed as a source. We frame this proposal as an error budget: quantum data offer advantage for generating potentials only if their total error, including basis set, statistical, and hardware errors, is smaller than that of the best classical method affordable for the same system, and only if enough data can be produced. We review the main quantum algorithms and recent demonstrations of potentials trained on quantum data in this light. The machine learning side of the workflow is ready: sparse, noisy data combined with transfer learning yield usable potentials. The electronic structure side - quantum data produced by quantum hardware - is not yet competitive, since current hardware demonstrations treat systems that classical methods already solve virtually exactly, in basis sets too small to be useful at this time. Nevertheless, we write this commentary so that researchers can evaluate the use of quantum data for MLIPs as quantum hardware continues to improve, hopefully to the point of advantage.

Type